Question about roi_head/num_bg_samples

[amida47]

in the logged losses we have a loss called roi_head/num_bg_samples I am wondering if it represents the number of background samples per image or per batch

because I am doubting that MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE sets the number of samples for a batch and not a image

example
I set
cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 10
cfg.SOLVER.IMS_PER_BATCH = 32

then on tensorboard

if we agree that these values are logged every batch, then
as you can see if we sum num_bg_samples and num_fg_samples , we will find a value equal to ROI_HEADS.BATCH_SIZE_PER_IMAGE

[ppwwyyxx]

I am wondering if it represents the number of background samples per image or per batch

They represent the average number of fg/bg samples per image.